AI-driven resource allocation in cloud computing: a systematic review revealing critical sustainability and evaluation gaps

Abstract This systematic literature review analyzes AI-driven resource allocation in cloud computing through comprehensive analysis of 63 high-quality studies selected via PRISMA 2020 methodology from an initial collection of 485 papers. Our taxonomic framework categorizes approaches across four dimensions: algorithmic methods, deployment environments, optimization objectives, and evaluation methods. Quantitative analysis demonstrates substantial AI superiority over traditional approaches: $$\approx $$ 45% average latency reduction (range 11−77.7%) from 10 studies with quantifiable latency data, 32% cost savings (range 10–48%) from 6 studies with quantifiable cost data, and 35% energy efficiency improvements (range 3.68–71%) from 16 studies with energy measurements. Reinforcement learning dominates the field (40% of studies) with particular effectiveness in dynamic environments, while hybrid approaches demonstrate superior multi-objective optimization. Critical research gaps include minimal carbon-aware scheduling integration (only 4 studies, 6.3% of corpus), over-reliance on simulation environments (70% of evaluations), and absence of standardized evaluation frameworks. The limited availability of quantifiable performance data across studies reveals a significant methodological gap in current research evaluation practices. We identify five high-priority research directions and provide actionable recommendations for advancing production-ready AI-driven cloud resource management systems that balance performance, sustainability, and practical deployment requirements.